pixelscribe

PixelScribe is a tool that enables multimodal large language models without native image generation capabilities to produce pixel art through a closed loop of PixelScript drawing language, visual feedback, and self-correction. The model writes PixelScript, an engine renders it in seconds, the resulting PNG is fed back to the model, the model self-evaluates, and then revises the script iteratively. The project states that the model lacks a hand rather than a brain, and PixelScribe provides that hand. It uses a compact DSL of 43 instructions, one per line, which the model can write correctly in one pass. The rendering engine supports procedural realistic rendering with lighting, relief, specular highlights, bloom, tone mapping, and fractal materials, and can optionally call neural backends for diffusion or img2img when configured. The tool includes a professional editor with layers, undo stack, symmetry drawing, palettes, flood fill, dithering, and outlining, and it runs with zero npm dependencies on Node 20+.